1972 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 53, NO.

8, SEPTEMBER 2008

Containment Control in Mobile Networks strategy is feasible in the sense of non-Zeno, live in the sense of not
staying in the Stop mode indefinitely, and convergent in the sense that
M. Ji, G. Ferrari-Trecate, M. Egerstedt, and A. Buffa the target area is in fact reached. This approach can also be generalized
to hierarchial networks, as was illustrated by our preliminary work in
[10].
Abstract—In this paper, the problem of driving a collection of mobile
robots to a given target destination is studied. In particular, we are inter-
ested in achieving this transfer in an orderly manner so as to ensure that the II. BACKGROUND AND MATHEMATICAL PRELIMINARIES
agents remain in the convex polytope spanned by the leader-agents, while
the remaining agents, only employ local interaction rules. To this aim we
In this section we will present the basic mathematical framework and
exploit the theory of partial difference equations and propose hybrid con- some enabling results in multi-agent control.
trol schemes based on stop-go rules for the leader-agents. Non-Zenoness, We start with basic notions of graph theory. For more details we refer
liveness and convergence of the resulting system are also analyzed. the reader to [11]. An undirected graph G is defined by a set NG =
Index Terms—Containment problems, decentralized control, graph 1 ...
f ; N g of nodes and a set EG  NG 2 NG of edges. We will also
theory, leader-following, multi-agent systems, partial difference equations. use jNG j for denoting the cardinality of NG . Two nodes x and y are
( )
neighbors if x; y 2 EG . The neighboring relation is indicated with
x  y and P x ()= :
fy 2 NG y  xg collects all neighbors to
I. INTRODUCTION the node x. A path x0 x1 ...
xL is a finite sequence of nodes such that

This paper investigates a particular subarea of multi-agent control, xi01  xi , i ; = 1 ...

; L. A graph G is connected if there is a path

namely the so-called containment problem where a collection of au- connecting every pair of distinct nodes.
tonomous, mobile agents are to be driven to a given target location Definition 1: Let S = ( )
NS ; ES be an undirected host graph
while guaranteeing that their motion satisfies certain geometric con- and NS  NS . The subgraph S 0 associated with NS is the pair
straints. These constraints are there to ensure that the agents are con- ( ) = ( ) :
NS ; ES where ES f x; y 2 ES x 2 NS ; y 2 NS g
tained in a particular area during their transportation. Such issues arise Definition 1 allows basic operations in set theory to be extended to
for example when a collection of autonomous robots are to secure and graphs. For instance, if S1 and S2 are two subgraphs of the graph S ,
then remove hazardous materials. This removal must be secure in the then S1 [S2 , S1 \S2 , S1 nS2 are the graphs associated with NS [NS ,
sense that the robots should not venture into populated areas or in other NS \ NS , and NS nNS , respectively. For our purposes, we will
ways contaminate their surroundings. often use graphs with a boundary.
We approach this problem from a leader-follower point-of-view Definition 2: Let S be a subgraph of G. The boundary of S is the
[1]–[3]. In particular, we will let the agents move autonomously based subgraph @ S  G associated with N@S
:
=
fy 2 N G n N S 9 x 2 :
on local, consensus-like interaction rules, commonly found in the liter- :
NS x  yg. The closure of S is S @ S [ S . =
ature under the banner of algebraic graph theory [4]–[6]. However, we Note that the definition of the boundary of a graph depends upon the
host graph G. This implies that if one considers three graphs S 0  S 
will augment this control structure with the addition of leader-agents
or anchor nodes [7]. These leaders are to define vertices in a convex 0
G, the boundaries of S in S and in G may differ.
polytope (the leader-polytope) and they are to move in such a way that
In the context of multi-agent systems, the nodes of the host graph G
the target area is reached while ensuring that the follower-agents stay
represent agents and the edges are communication links. In particular,
in the convex polytope spanned by the leaders, up to a given tolerance.
As such, the followers movements are calculated in a decentralized an agent x has access to the states of all its neighbors and can use this
manner according to a fixed interaction topology, while the leaders piece of information to compute its control law. Although a complete
are assumed to be able to detect if any of the followers violate the graph is not necessary for a distributed control algorithm, we always
containment property. assume that the host graph is connected.
For the leaders, we will use a hybrid Stop-Go policy [8], [9], in In order to model the collective behavior of the agents we will use
which the leaders move according to a decentralized formation control functions f :
NG 7! d defined over a graph G[12]. The partial
strategy until the containment property is about to be violated. At this derivative of f is defined as @y f x
:
( )= ( )
f y 0 f x and the Laplacian ()
point, they stop and let the followers settle back into the leader-polytope of f is given by
before they start moving again. For such a strategy to be successful,
a number of results are needed, including a guarantee that the Lapla- 1 ( )=0
f x
: 2
( )=+
@y f x ()
@y f x ; (1)
cian-based follower-control will in fact drive the followers back into y 2N ;y x y 2N ;y x
the leader-polytope. Moreover, we must also ensure that such a control
where the last identity follows from the fact that @y2 f x 0 @y f ( )= ( ).
x

The integral and the average of f are defined, respectively, as

Manuscript received August 24, 2006; revised September 3, 2007. Current
version published September 24, 2008. The work by G. Ferrari-Trecate was par-
tially supported by the European Commission under the Network of Excellence f dx =
:
f x ;( ) h i = jN1 j
f
:
f dx: (2)
HYCON, contract number FP6-IST-511368. The work by M. Egerstedt and G x2N G G

M. Ji was supported by the U.S. Army Research Office through Grant #99838.
Recommended by Associate Editor J. Hespanha. Let L
2
( j ) be the Hilbert space composed
G
d
by all functions : f

N 7! endowed with the norm k k2 = k k2 . We will use the

M. Ji and M. Egerstedt are with the Georgia Institute of Technology, School d
G f f
of Electrical and Computer Engineering, Atlanta, GA 30332 USA (e-mail: L G
2
[email protected]; [email protected]). shorthand notation L when there is no ambiguity on the underlying
G. Ferrari-Trecate is with the Dipartimento di Informatica e Sistemistica, Uni- domain and range of the functions.
versità degli Studi di Pavia, 27100 Pavia, Italy and also with INRIA, Domaine de Let S be a subgraph of G and @ S be its boundary in G. We as-
Voluceau, Rocquencourt—B.P.105, 78153, Le Chesnay Cedex, France (e-mail: =
sume that S [ @ S G. As in [12], we also consider the Hilbert space
[email protected]). 1
H0 S ( )= f 2 L
2
G fj@S ( ): =0
(see [12] for the definition of a
( )
suitable norm on H01 S ). Note that a function f 2 H01 S is defined ( )
A. Buffa is with the Istituto di Matematica Applicata e Tecnologie Infor-


matiche, C.N.R., 27100 Pavia, Italy (e-mail: [email protected]).
Digital Object Identifier 10.1109/TAC.2008.930098 on S and possibly non null only on S .

0018-9286/$25.00 © 2008 IEEE

Authorized licensed use limited to: UNIVERSITY OF PAVIA. Downloaded on December 19, 2008 at 12:05 from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 53, NO. 8, SEPTEMBER 2008 1973

The next theorem, proved in [12], characterizes the eigenstructure of and they have been studied in [12]. In particular, [12, Theorem 3.5]
the Laplacian operator defined on H01 S . () =
shows that if G is connected and NS 6 ; then, the PdE (5) has a
Theorem 1: Let G be a connected graph and S a proper subgraph ()
unique solution2h x . By analogy with the jargon of Partial Differen-
of G. Then, the operator 1: ( ) ( )
H01 S j d 7! L2 S j d has jNS jd tial Equations, h is termed the harmonic extension of the boundary con-
strictly negative eigenvalues. Moreover, the corresponding eigenfunc- 
ditions r .
( )
tions form a basis for H01 S j d . Our next aim is to verify that r ! h as t ! 1. Let us consider+
the decomposition
III. MULTIPLE STATIONARY LEADERS
r(x; t) = r0 (x; t) + h(x); r0 2 H01 (SF ): (6)
In this section, we use PdEs for modelling and analyzing a group
of agents with multiple leaders. A leader is just an agent that moves Since h does not depend upon time and 1h = 0 , 8 x 2 NS , the PdE
toward a predefined goal, and whose control policy is independent of (4) is equivalent to the following one
the motion of all the followers. However, followers that are neighbors
to the leader can use the leader state in order to compute their control r_0 (x; t) = 1r0 (x; t) x 2 NS (7a)
inputs. r0 (x; t) = 0 x 2 NS : (7b)
( )
Let r x; t be the position of the agent x at time t  , where1r 2 0
L2 . The communication network is represented by the undirected and From (6), it is apparent that the problem of checking if r ! h as
connected graph G. For distinguishing between leaders and followers, t ! +1 can be recast into the problem of studying the convergence to
we consider two subgraphs SF and SL of G and assume that SL = zero of the solutions to the PdE (7). The fact that r0 ! 0 as t ! +1
=
@SF and SF [ SL G, where the subscripts denote ”Leaders” and follows from Theorem 1 and it can be shown by proceeding exactly as
”Followers” respectively. Note that we assume that all agents are either in the proof of [17, Theorem 5]3.
designated as leaders or followers. The next Theorem, proved in [19], highlights a key geometrical fea-
As already mentioned in the introduction, we will assume that the () ( )
ture of h x . For a set X of points in d , Co X will denote its convex
_( ) = ( )
followers obey the simple dynamics r x; t u x; t , where hull. Moreover, the set L is the convex hull of leaders positions, i.e.,
: Co fr y ; y 2 N g .
:
u(x; t) = 1r(x; t)
L = ( ( ) S )
(3) Theorem 2: Let S1 be a nonempty connected subgraph of SF and
@S1 be its boundary in G. Then, 8 x 2 NS it holds
is the Laplacian control law. Let r^(x; t), x S be the trajectory of
2 N

the leaders. Then, the collective dynamics is represented by the model h(x) 2 Co(fh(y); y 2 N@S g ): (8)
r_ (x; t) = 1r(x; t) x 2 NS (4a) ()
Moreover, one has that h x 2 L , i.e., that the position of each fol-
r(x; t) = r^(x; t) x 2 NS (4b) lower lies in the convex hull of the leaders positions. Finally, if L is
()
full-dimensional4, then h x 2 L n@ L , 8 x 2 NS .
endowed with the initial conditions r 1; ( 0) = ~
r 2 L2 SF . ( ) Another geometrical feature which we need is the following:
Model (4) is an example of a continuous-time Partial difference Theorem 3: Suppose that L is full-dimensional and that r x; t is ( )
Equation (PdE) with non-homogeneous Dirichlet boundary conditions. evolving according to (4). Suppose that, at a given time t t, there =
We refer the reader to [12]–[14] for an introduction to PdEs. ( )
is an agent x 2 NS such that r x; t 2 @ L and r y; t 2 L , ( )
The main results on Laplacian control available in the literature and ()
8 y 2 P x . Then, two situations may occur:
specialized to model (4) are: 1) there exists an (affine) hyperplane  such that
=
• in the leaderless case (i.e., SL ;), the Laplacian control solves
( )
the rendezvous problem, i.e., r x; t ! r3 2 d ; 8 x 2 NG as r(x; t) 2  \ @ L; and r(y; t) 2 \@ L 8 y 2 P (x ):
+
t ! 1. Moreover, the agents converge exponentially to r3 =
~
hr i thus achieving average consensus. These results have been Then

> 0 : r(x; t) + r_ (x; t) 2  \ @

established in [15], [16] through the joint use of tools in control
theory and algebraic graph theory. A formal analysis of the PdE 9 L (9)
(4a) has been conducted in [13], [14], [17] showing a complete
accordance with results available within the theory of the heat 2) otherwise
equation [18];
• in the case of a single leader (i.e., NS =
fxL g) with fixed posi-
> 0 : r(x; t) + r_ (x; t) 2 @
^(
tion (i.e., r xL ; t)= r 2 d ), Laplacian control solves the ren- 9 L n L: (10)
dezvous problem with r3 =
r [15]. This property has also been
Note that (9) means that the velocity of x will be along the hyperplane
shown in [13], [14] within the PdE framework, thus highlighting
the profound links between model (4) and the heat equation with  (in other words, the agent may slide on the boundary @ L ), whereas
(10) means that the velocity of x is pointing inside the polytope L .
While Theorem 2 and the fact that r ! h as t ! +1 guarantee
Dirichlet boundary conditions [18].
The first attempt of this paper is to characterize the asymptotic be-
havior of the followers in the presence of multiple leaders with fixed that followers asymptotically enter L , Theorem 3 ensures that once
positions. To this end, for the remainder of this section, we will assume all followers are in L they cannot exit from this set and therefore
^( ) = ( )
that r x; t ( )
r x 2 L2 SL . The equilibria of (4) are then given by containment will be never violated.
the solutions to the PdE 2[12, Theorem 3.5] assumes that the subgraph is induced (see [12] for the
definition of induced subgraphs). However, a careful examination of the proof,
reveals that this assumption is unnecessary.
1 h (x ) = 0 x 2 NS (5a) 3Actually, [17, Theorem 5] proves a stronger property, namely that the origin

h(x) = r(x) x 2 NS (5b) of (7) is “exponentially stable on the space ( )

”. The definition of stability
of equilibria on subspaces is provided in [17].
1For sake of conciseness, for a function ( ): we 4The set is full-dimensional if the dimension of the affine hull
will often write instead of ( ) . generated by is (see [20]).

Authorized licensed use limited to: UNIVERSITY OF PAVIA. Downloaded on December 19, 2008 at 12:05 from IEEE Xplore. Restrictions apply.
1974 IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 53, NO. 8, SEPTEMBER 2008

Proof: (Theorem 3): Since r(x; t) obeys to (4), by rearranging only need access to the positions (and target locations) of their neigh-
terms we obtain boring leaders in order to compute their motion. As such the decentral-

r_ (x; t) = 0jP (x)jr(x; t) + r(y; t):

ized character of the algorithm is maintained.
In order to fully specify the hybrid Stop-Go leader policy transition
y 2P (x) rules are needed as well. As before, let L denote the leader-polytope
Then, setting = jP (x)j01 , it holds ( )
and let d ; L denote the signed distance

r(x; t) + r_ (x; t) = jP (x)j01 r(y; t);

d(; L ) =:  ()x2min
@
k 0 x k 2 ; (13)

y 2P (x) where k 1 k2 denotes the Euclidean 2-norm, and where   0 ( )= 1

:
i.e., r(x; t) + r_ (x; t) is the barycenter b(Yx ) of the polytope Yx =
if  2 L and +1otherwise. Using this distance measure we let the

Co(fr(y; t); y 2 P (x)g). Note that, if r(y; t) 2 L , 8 y 2 P (x) two guards, i.e., transition conditions, be given by
one has Yx 2 L . Moreover, thanks to convexity, the barycenter of GO 2ST OP : 9y 2 NS j d(r(y; t); )  0?
Yx lies in the relative interior of Yx . Thus, if all y 2 P x verify that () ST OP 2GO : d(r(y; t); L ) < 0
L

8 y 2 NS ?
(14a)

( )
r y; t 2  \ @ L then Yx   \ @ L and so does b Yx , i.e., ( )
( ) ( )
b Yx 2  \ @ L ; otherwise b Yx 2 L n @ L . (14b)

where a transition from GO to STOP triggers when the conditions in

IV. LEADER-FOLLOWER CONTAINMENT CONTROL
GO2STOP are met, and similarly for STOPT2GO , and where  > is 0
Containment of all the followers is achieved in the case of static a threshold.
leaders in the last section. However, if the leaders are moving, this prop- Note that the guard STOP2GO is crossed only if the following as-
erty might be violated. In order to prevent the followers from leaving sumptions are verified:
the polytope spanned by the leaders, appropriate control strategies need ^( )
Assumption 1: Let h 1; t be the solution to (5) for r 1 ( ) = ^( )
r 1; t ,
to be designed for the leaders to guarantee the containment. In what fol- 0 ()=
8 t  and consider the set L t fy 2 L t d y; @ L t < (): ( ( ))
lows, we propose a hybrid strategy for this purpose and analyze liveness 0g. Then
and reachability of the resluting closed-loop system. 1) L 
()
t is nonempty, 8 t  ; 0
( ^( )
2) Co fh x; t ; x 2 NS g  L 
)t. ()
A. Hybrid Control Strategy 0 ^( )
Note that, for a given time t  , the uniqueness of h 1; t follows from
For the sake of containment, we define two distinctly different con- the uniqueness of the solution to (5). In particular, Assumption 1 im-
trol modes for the evolution of the leaders. The first of the two control plies that L must be full-dimensional at all times and “sufficiently fat”
modes is the STOP mode that corresponds to the leaders halting their along every direction (see condition 1). Conditions relating property 2
movements altogether in order to prohibit a break in the containment: of Assumption 1 to the graph topology are currently under investiga-
tion. A few comments must be made about the computation and com-
ST OP : (4a); (4b) and r^_ (x; t) = 0; x 2 NS : (11) munication requirements that these guards give rise to. If two leaders
are located at the end-points of the same face of L , then they must be
It is clear that in order to execute this mode, no information is needed able to determine if any of the followers are in fact on this face. This can
for the leaders whatsoever. be achieved through a number of range sensing devices, such as ultra-
The second control mode under consideration is the GO mode, in sonic, infra-red, or laser-based range-sensors. Moreover, in order for all
which the leaders move toward a given target formation. A number leaders to transition between modes in unison, they must communicate
of different control laws can be defined for this, but, for the sake of between them, which means that either SL is a complete graph, or that
conceptual unification, we let the GO mode be given by a Laplacian- multi-hop strategies are needed. In either way, a minimal requirement
based control strategy as well. for these mode transitions to be able to occur synchronously, without
having to rely on information flow across follower-agents, is that SL
GO : (4a); (4b) and r^_ (x; t) = 1S (^r(x; t) 0 rT (x)); must be connected.
x 2 NS (12) 0
The hysteresis threshold  > in the STOP2GO guard and the next
assumption are needed in order to avoid Zeno behaviors. Let  de-
()
where rT x ; x 2 NS denotes the desired target position of leader x note the supremum of the diameter of L during an execution.
1
and S denotes the Laplacian operator defined solely over the sub- Assumption 2: 9M < 1 such that   M.
graph SL , i.e., It is easy to check that Assumption 2 is verified when Laplacian con-

1S f (x) =: 0
trol governs the leaders’ motion in the GO mode as in (12). Indeed, the
@y2 f (x): ^( )
exponential convergence of r x; t to rL x ( ) = ^( 0)
hr 1; 0 rT 1 i () +

y x; y 2N () ^( )
rT x implies that r x; t is bounded at all times. However, Laplacian
control is but one of many possible control strategies and can be re-
Under the assumption that SL is connected, and by exactly the same placed by other control schemes (e.g., plan-based leader control laws)
reasoning as for the standard rendezvous problem, under the influ- without generating Zeno executions as long as Assumption 2 is verified.
ence of the GO mode alone the leaders will converge exponentially
( ) = ^( 0)
to rL x () + ( )
hr 1; 0 rT 1 i rT x , i.e., 9k > ;  > such that 0 0 Theorem 4: Under Assumptions 2 and 1, the hybrid automaton de-

kr 1; t 0 rL x kL  ke0t kr 1; 0 rL x kL . In other words,

^( ) ( ) ^( 0) () fined by (11), (12) and (14) is non-Zeno.
Proof: Let the system be in the STOP mode. Under Assumption
no convergence to a predefined point is achieved. Rather, this control 2 we have
law ensures that the leaders arrive at a translationally invariant target
formation. kr_ (x; t)k = k1r(x; t)k  k@ y r (x )k
Note that the details of the leaders’ motion is not crucial and this par- 
y x
ticular choice is but one of many possibilities. However, this choice is    N ; 8 x 2 NS : (15)
appealing in that it makes the information flow explicit, and the leaders 
y x

Authorized licensed use limited to: UNIVERSITY OF PAVIA. Downloaded on December 19, 2008 at 12:05 from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON AUTOMATIC CONTROL, VOL. 53, NO. 8, SEPTEMBER 2008 1975

From Assumption 1, in order for the system to leave the STOP mode, at REFERENCES
least one follower agent must have travelled at least a distance , which
[1] J. Desai, J. P. Ostrowski, and V. Kumar, “Controlling formations of
in turn implies that the system will always stay for a time greater than or
multiple mobile robots,” in Proc. IEEE Int. Conf. Robot. Automat,
equal to =N  in the STOP mode. In order for the system to exhibit 1998, pp. 2864–2869.
Zeno executions, a necessary condition is that the difference between [2] M. Ji, A. Muhammad, and M. Egerstedt, “Leader-based multi-agent
the transition times must approach zero [21]. Since this is not the case coordination: Controllability and optimal control,” in Proc. American
here, the non-Zeno property is established. Control Conf., 2006.
[3] H. Tanner, G. Pappas, and V. Kumar, “Leader to formation stability,”
B. Liveness and Reachability IEEE Trans. Robotics Autom., vol. 20, no. 3, pp. 443–455, Jun. 2004.
[4] H. Tanner, A. Jadbabaie, and G. Pappas, “Flocking in fixed and
As already mentioned, the proposed solution is non-Zeno. However, switching networks,” IEEE Trans. Automat. Control, vol. 52, no. 5,
as it is currently defined, the Stop-Go policy may be blocking in the pp. 863–867, May 2007.
sense that the system never leaves the STOP mode. One remedy to this [5] R. O. Saber, A Unified Analytical Look at Reynolds Flocking Rules
problem is to allow the containment to be slightly less tight. In other California Institute of Technology, Pasadena, CA, 2003, Tech. Rep.
CIT-CDS 03-014.
words, we can select different guards, e.g.,
[6] A. Muhammad and M. Egerstedt, “Connectivity graphs as models of
2
GO ST OP : 9y 2 NS jd(r(t; y); L ) > 2? (16a) local interactions,” J. Appl. Math. Comput., no. 1, pp. 243–269, 2005.
[7] H. G. Tanner, “On the controllability of nearest neighbor interconnec-
ST OP 2GO : d(r (t; y ); L )   8 y 2 NS ? (16b) tions,” in Proc. 43rd IEEE Conf. Decision and Control, Dec. 2004, pp.
2467–2472.
where  > 0. What this means is that we do not enter the STOP mode [8] M. Egerstedt and C. Martin, “Conflict resolution for autonomous vehi-
until a follower is 2 outside L . Let us define cles: A case study in hierarchical control design,” Int. J. Hybrid Syst.,
vol. 2, no. 3, pp. 221–234, 2002.
L; = fy 2 : d(y; L )   g:
: d [9] H. Sussmann, “A maximum principle for hybrid optimal control prob-
lems,” in Proc 38th IEEE Conf. Decision and Control, 1999, vol. 1, pp.
Note that, one has L  L; . The next Theorem summarizes the 425–430.
main properties of the resulting hybrid automaton. A remarkable fea- [10] M. Ji, M. Egerstedt, G. Ferrari-Trecate, and A. Buffa, “Hierarchical
containment control in heterogeneous mobile networks,” in Proceed-
ture of the guards (16) is that Assumption 1 is no longer needed in order
ings Mathematical Theory of Networks and Systems. Kyoto, Japan,
to guarantee liveness. 2006, pp. 2227–2231.
Theorem 5: Under Assumption 2, the hybrid automaton by (11), [11] B. Bollobás, Modern Graph Theory, ser. Graduate texts in Mathe-
(12) and (16) is non-Zeno, live, in the sense of always leaving the matics. Berlin/New York: Springer-Verlag, 1998.
STOP mode eventually, and convergent in the sense that r^(x; t) ! [12] A. Bensoussan and J.-L. Menaldi, “Difference equations on
hr^(1; 0) 0 rT (1)i + rT (x). weighted graphs,” J. Convex Anal. (Special issue in honor of Claude
Lemaréchal), vol. 12, no. 1, pp. 13–44, 2005.
Proof: We first prove liveness. Assume that the system is in the
STOP mode. From Theorem 2 we have that h 2 L . Since 8 x 2 SF ,
[13] G. Ferrari-Trecate, A. Buffa, and M. Gati, Analysis of Coordination
in Multiple Agents Formations Through Partial Difference Equations
r (x; t) ! h, and L  L; , every follower will eventually get back IMATI-CNR, Italy, 2004, Tech. Rep. 5-PV [Online]. Available: http://
in L; in finite time (recall that the leaders are stationary in the STOP sisdin.unipv.it/lab/personale/pers_hp/ferrari/publications.html
mode) hence triggering a transition to the GO mode. [14] G. Ferrari-Trecate, A. Buffa, and M. Gati, “Analysis of coordination
Under Assumption 2, it holds kr_ (x; t)k  N ( + 2) and we in multi-agent systems through partial difference equations. Part I: The
Laplacian control,” in Proc. 16th IFAC World Cong. Automatic Control,
can repeat the non-Zeno argument in the proof of Theorem 4 in order 2005.
to see that the system always stays in the GO mode for a time greater [15] A. Jadbabaie, J. Lin, and A. S. Morse, “Coordination of groups of mo-
than or equal to =(N ( + 2 )). bile autonomous agents using nearest neighbor rules,” IEEE Trans. Au-
As a result, in a non-blocking system the leaders will be given infin- tomat. Control, vol. 48, no. 6, pp. 988–1001, Jun. 2003.
[16] R. Olfati-Saber and R. Murray, “Consensus problems in networks
itely many opportunities to move during a finite (bounded away from
of agents with switching topology and time-delays,” IEEE Trans.
zero) time horizon, which implies convergence to the target location as Automat. Control, vol. 49, no. 9, pp. 101–115, Sep. 2004.
long as the leaders would in fact end up at the target location under the [17] G. Ferrari-Trecate, A. Buffa, and M. Gati, “Analysis of coordination in
influence of the GO mode alone. multi-agent systems through partial difference equations,” IEEE Trans.
Autom. Control, vol. 51, no. 6, pp. 1058–1063, Jun. 2006.
V. CONCLUSIONS [18] R. Dautray and J.-L. Lions, Mathematical Analysis and Numerical
Methods for Science and Technology. Berlin: Springer-Verlag, 1992,
In this paper, we presented a hybrid Stop-Go control policy for the vol. 5–6, Evolution problems I–II.
leaders in a multi-agent containment scenario. In particular, the control [19] M. Ji, G. Ferrari-Trecate, M. Egerstedt, and A. Buffa, Contain-
strategy allows us to transport a collection of follower-agents to a target ment Control in Mobile Networks Georgia Institute of Tech-
area while ensuring that they stay in the convex polytope spanned by nology, 2007, Tech. Rep. GT-GRITS-07-01 [Online]. Available:
the leaders. The enabling results needed in order to achieve this is that, http://www.ece.gatech.edu/~magnus/GT-GRITS-07-01.pdf
[20] S. Boyd and L. Vandenberghe, Convex Optimization, Cambridge Univ.
for stationary leaders, the followers in a connected interaction graph
Press, Cambridge, U.K., 2004.
will always converge to locations in the leader-polytope. Extensions [21] K. Johansson, M. Egerstedt, J. Lygeros, and S. Sastry, “Regulariza-
to the proposed control strategy are moreover given in order to ensure tion of zeno hybrid automata,” Syst. Control Lett., vol. 38, no. 6, pp.
certain liveness properties. 141–150, 1999.

Authorized licensed use limited to: UNIVERSITY OF PAVIA. Downloaded on December 19, 2008 at 12:05 from IEEE Xplore. Restrictions apply.